DETECTION OF PATRONIZING AND CONDESCENDING LANGUAGE (PCL) USING A TRANSFORMER-BASED MODEL

• Main Author: Abraham Sianturi, Gerald
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85073
• Institution: Institut Teknologi Bandung

Patronizing and condescending language (PCL), although often used with good intentions, can lead to discrimination, perpetuate negative stigma, and hinder the inclusion of vulnerable groups. The detection of patronizing and condescending language was highlighted in SemEval 2022 as the fourth task, consisting of binary classification as well as multilabel PCL classification. This study discusses the development of a PCL detection model using a transformer-based approach due to its ability to capture complex linguistic features. Experimental results show that in the first task, binary classification, the DeBERTa-v3-large model achieved a better f1- score performance than LLaMA-3-8B, with scores of 0.549 and 0.361, respectively. Data augmentation yielded varying performance depending on the subtask and model used, indicating inconsistency in its application. Moreover, using binary relevance, the weighted random sampling strategy effectively enhanced performance on the multilabel task. The performance results using task transformation strategies, such as binary relevance and label powerset, produced macro-average f1-scores of 0.346 and 0.17, respectively, which were unable to surpass the model performance achieved in SemEval 2022 Task 4.